[LockerGnome] I can't believe I rebooted the tech news & tips site!
Chris Pirillo's LockerGnome is a reader-supported publication. To receive new posts and support my work, consider becoming a free or paid subscriber.
Greetings, Programs!
First off, the LockerGnome reboot is going like gangbusters! So many fresh (and fun) articles are being posted every day. Best to get the RSS feed for that, because there’s no way I can include all that in an email digest - and link roundups may cause click blindness.
In this week’s newsletter, I asked Android users what's stopping them from getting an iPhone, priced the $999 iPhone Duo inner screen repair, and read the Surface Laptop Ultra reviews. There's also a Commodore 64 you can break pin by pin, a pond of synthesized ducks, and Windows 98 getting USB 3 drivers 28 years late.
Thanks to 10 of you for supporting the reboot…
I asked Android users what’s stopping them from getting an iPhone…
I asked Android users in a YouTube poll on October 2 what’s stopping them from getting an iPhone, and most of the people who answered just like Android more. That was 61% of the first 93 votes, and another 18% said they might switch. On Facebook, a screenshot I shared of the same question has been seen 1.25 million times, with nearly 5,000 comments.
In the March quarter, 12% of new U.S. iPhone buyers came over from an Android phone, according to CIRP’s own report, down from 14% a year earlier, as ChannelNews reported it. CIRP put iPhone loyalty at 87% for the same quarter, MacRumors reported.
One of Android’s big differences is getting harder to use. Google’s new developer verification rules went live on September 30 in Brazil, Indonesia, Singapore and Thailand, and GSMArena reported that installing an app from an unverified developer there now means turning on a Developer Options setting, authenticating, restarting the phone and waiting 24 hours. After the wait, the setting lasts seven days or indefinitely, as the user chooses. Google says it will expand those protections globally in 2027. If installing anything you like is why you stay, I’d watch that year.
Moving is easier than it used to be, and you still leave things behind. Apple’s support page says Move to iOS brings over contacts, message history, photos, calendars, call history and WhatsApp messages. Music, books and PDFs have to be moved by hand. Apple’s guide doesn’t mention the apps you paid for, so I’d plan on buying those again.
If you’re on the fence, I’d make two lists before buying. The first is every app and subscription you’ve bought on your current phone. The second is the people you text most and which phones they use.
If both lists say switch, I’d buy the iPhone from a store with a return window and keep the old phone in a drawer until the window closes.
Meta opened up Muse Spark, and it takes two Mac Studios to run it at home…
Meta released Muse Spark Open on October 15, a model with 816 billion parameters, 42 billion of them active. MacRumors reports it’s free to download under a license that allows limited commercial use. Meta calls it the most capable U.S. open-weight model to date, and says you can run it on your desktop using two Apple M5 Ultra Mac Studios.
MacObserver reported Apple and Meta teamed up to run it on the Mac Studio, and Apple’s machine learning research team published the numbers. Meta says two are enough because each M5 Ultra Mac Studio’s GPU can use up to 512 GB, and in a 4-bit format the model takes about 493GB, Apple wrote. The Macs talk over a low-latency driver on Thunderbolt 5, and Apple’s speed figure, over 36 tokens per second, comes from four of them. That’s Apple’s test, so I’d wait for someone else to time it.
The Decoder reported the license requires a “Built with Muse Spark Open” label. Digital Applied reported that model-as-a-service providers above $20 million in revenue must notify Meta and sign a separate agreement, while Trending Topics reported that threshold doesn’t apply to internal use, research or building the model into your own products. Even the download is gated: the Hugging Face page asks you to log in and review the conditions first.
It isn’t Meta’s best model, either. The Hugging Face model card lists a score of 39 on the Artificial Analysis Intelligence Index, and The Decoder says it still falls well short of Meta’s paid Muse Wonder Full models.
If you’re curious, I wouldn’t buy two Mac Studios for this. Meta says it’s already available through Baseten, Fireworks AI and Groq, so I’d try it there first, read the license before building anything a customer pays for, and watch for smaller versions that fit one machine. If you already own two M5 Ultra Mac Studios, you’re the audience Apple had in mind.
Open weights are the headline. Two Mac Studios are the fine print.
Withings wants about $100 a year to sync your scale with Apple Health…
Anyone who creates a Withings account on October 13 or later needs a Withings+ subscription for automatic syncing to Apple Health, MacRumors reported, at $99.50 a year in the U.S. Tech Between the Lines found the App Store listing at $9.99 a month or $99.99 a year. Accounts created on or before October 12 keep syncing for free.
iThinkDiff reported that a new account on Android also needs it for automatic transfers to Health Connect. A reply marked as Withings’ official response, quoted by Engadget, says “automatic sync to third-party apps is part of Withings+.”
Gadgets & Wearables reported that Withings says subscription revenue helps fund clinical validation, regulatory work and the infrastructure behind it, and iThinkDiff reported Withings ties the change to API updates and new privacy measures for its AI connector. Either is a fair reason to charge for a service on top of the device.
The rule also covers other apps that pull data through Withings’ free developer plan, including apps already on the market, iDevice reported, and new users can’t get around it by routing their data through Apple Health or Google’s system instead. A thread on Withings’ own community calls the change unacceptable right in its title.
Before I bought any connected scale, blood pressure cuff or watch, I’d check three things: whether syncing to Apple Health or Google’s health app needs a paid plan, whether I can export my full history as a file, and what happens to that data if I stop paying. Withings still lets every user export a full history as a CSV or PDF, iThinkDiff reported. If you already own a Withings device, I’d hang on to that account, because the exemption follows the account, not the device.
My weight, sent from my own bathroom floor to my own phone, is a strange thing to rent back to me.
iPhone Duo inner screen repair costs $999, so if you preorder one anyway, buy AppleCare+…
iPhone Duo preorders opened October 16, and the number I’d look at before the $1,999 starting price is $999. That’s Apple’s estimated fee to repair the Duo’s folding inner screen without AppleCare+, MacRumors reported from Apple’s own repair listing, and iDevice reported the same figure. With AppleCare+, the same inner display repair is $99, according to 9to5Mac. Apple says the phones start arriving October 23.
The rest of Apple’s list, as MacRumors reported it, is $379 for the outer display, $199 for back glass and up to $1,499 for anything filed under other damage. iThinkDiff notes these are estimates, and the final fee is set after Apple inspects the phone.
Samsung’s folding phones cost less to fix. Nokiamob reported that the inner screen on a standard Galaxy Z Fold 8 runs $579, up from $449 for the Fold 7. Apple’s fold costs $420 more than the standard Fold 8 to replace.
9to5Mac reported AppleCare+ adds $19.99 a month to the Duo, and iThinkDiff reported you can also pay $199.99 a year or $399 up front. Paid yearly, the plan costs about a fifth of one inner-screen repair, so it pays for itself the first time the fold breaks.
In my October 9 YouTube video on why I won’t buy a Duo, I said that if you get one anyway, get protection, because if you’re dropping $2,000 you’re not going to complain about another few hundred. There’s no track record for this screen yet because the phone is brand new.
If you’re preordering anyway, AppleCare+ belongs in the same checkout, not on a list of things to decide later. If $199.99 a year on top of $1,999 feels like too much, I’d take that as the answer on the first fold and look at a phone without a hinge, or wait for the second Duo.
The inner screen is the part you open and close all day, and one cracked fold costs about half of what the phone did.
Microsoft’s Surface Laptop Ultra costs up to nearly $6,000, and the reviews are in…
Microsoft’s Surface Laptop Ultra finally has a price. It starts at $2,599 and the flagship configuration costs $5,899.99, Gadget Review reported, with preorders that opened October 7 and units available from October 16. Gizmodo noted the base model has 24GB of memory and a lower-end CPU and GPU, so the $2,599 machine isn’t the one it reviewed, a nearly $6,000 unit with the maximum 128GB.
Microsoft lists up to 128GB of unified memory shared between the CPU and GPU, which is what lets a laptop hold a big AI model on its own. Tom’s Hardware’s verdict fits on one line: Nvidia’s Windows on Arm platform impresses in local AI, and gaming isn’t its strong suit.
The MacBook Pro still wins the straight race. Gizmodo measured the M5 Max close to 44% ahead in Geekbench 7’s multi-core test, Macworld pointed out that the M5 Max has more than twice the memory bandwidth, and The Short Cut said the MacBook Pro is still the better option for most people.
This is an Arm chip, and Gizmodo reported that anything without an Arm version leans on Microsoft’s Prism emulator. CGMagOnline found games work only when they don’t rely on unsupported anti-cheat software, and The Short Cut couldn’t run most games at a playable frame rate at full resolution, even with DLSS. Battery depends on the job: CGMagOnline drained it in three to four hours under heavy work, while The Short Cut got through eight- and 10-hour days.
If you’re tempted, I’d list the five programs you can’t work without, plus any game or plug-in with anti-cheat or a driver, and check each one against Windows on Arm before paying. I’d buy the memory for the models you’ll actually run, not the biggest number on the page. And I’d know the return window before the box shows up.
A developer pulled its horror drinking game’s demo, afraid AI vibe coders would clone it…
Alliance Arts was going to release a public Steam demo of Drunken Goddess Reflux on October 19, and then it canceled it. PC Gamer describes the game as a “shot glass Russian roulette” drinking game against an anime demon girl, and the studio said on X it will hold closed online play sessions for a limited number of participants instead.
The reason is what got my attention. As PC Gamer quoted the studio, “We’re concerned that the demo’s content could be used to quickly imitate or replicate the game’s presentation and gameplay with AI or other tools, and that such works could be released before launch.” Twisted Voxel reported the team also concluded it lacked the resources to enforce the demo’s rules on posting and streaming gameplay footage.
The fear has some history behind it. PC Gamer pointed to a developer who shared her game prototype online in March and had it cloned within five hours, and to McPixel developer Sos Sosowski, who has shared examples on Bluesky of indie games copied almost instantly by AI cloners and released on Steam.
Kotaku’s Ethan Gach had a drier read. He wondered whether the studio simply decided Steam Next Fest was too crowded with junk demos and that it was easier to stand out by not offering one (“In that case, mission accomplished”), and he noted the game’s Steam page already had over 25,000 followers.
If you still want to play it, KeenGamer reports the playtest codes are handed out by lottery through the official Discord, with applications open until October 22 at 11:59 PM Japan time. Twisted Voxel says the game launches December 8 on Steam for Windows. If you make games yourself, PC Gamer’s point cuts both ways: sharing an idea early is one of the best ways to build a community, and it’s also the moment a copycat sees it.
Now even the demo is a secret.
Regex Marble Run turns your regular expression into a 3D marble run…
Regex Marble Run is a free web page that builds your regular expression as a 3D machine. You type a pattern and a test string, and each letter becomes a marble that rides to the gate that tests it. A miss falls into Reject, and a match rolls to the cup. The person who posted it to Hacker News, where it drew 85 points, said a coworker who used to teach non-technical people to code built it because students always found regex hard.
Every part of the pattern gets its own piece of hardware. A brass arch reads one character, a diverter chooses a branch for the pipe symbol, greedy loops try another round first while lazy loops try the exit first, and a lookahead sends a small marble down a side track to report back. You can press Play or step through one test at a time.
The site says it lets you watch catastrophic backtracking happen, and every developer should see that once. Stack Overflow’s postmortem says a post with roughly 20,000 consecutive characters of whitespace hit a backtracking regex and caused a 34-minute outage on July 20, 2016. Cloudflare went down for 27 minutes on July 2, 2019, after a single WAF rule with a poorly written regular expression exhausted its CPUs, and at one point it had lost 80% of its traffic.
RE2, Go and Rust use linear-time engines that never revisit a position, so the same pattern that floods this machine runs in time proportional to the string length there.
If you write regex for anything that takes user input, I’d paste in your scariest pattern with a long test string and watch the marbles pile up. One of its built-in examples is a long run of spaces between two characters, which is close to what bit Stack Overflow. The poster disclosed it was built with Agent Duel, a coding-agent tool their company makes.
Watch the marbles before your users do.
Someone simulated a Commodore 64 pin by pin, so you can break it on purpose…
A developer who goes by antirez put c64-hw-sim on GitHub, and it simulates the Commodore 64’s motherboard (the ASSY 250407 board) pin by pin. You can lift the leg of a chip, short it to ground or to +5V, pull a chip out of its socket or open a resistor, then watch what the machine does. Hackaday’s Tyler August wrote that you can do all of that in software, with no risk of harming antique hardware.
The part I love is that the program has no idea what a broken C64 is supposed to look like. Its README says it knows the board’s netlist and how each chip behaves at its pins, so when the screen fills with garbage, that garbage was computed from the circuit. In its own example, pulling the character ROM while the classic 10 PRINT maze runs leaves the CPU, the RAM and the program working, and only the shapes of the characters are lost.
It’s useful for repair, too. The README says that if you have a broken C64 and a hunch about what failed, you can simulate it first and desolder as few chips as possible, though it warns that different faults can look the same. On Hacker News, where antirez’s post drew 227 points, one commenter said it may come in handy when diagnosing and fixing C64 boards.
It’s slow (every pin of every chip is simulated, after all). The README lists 31% of real speed on an Apple M3 Max and 48% on an M5 Max. Hackaday notes it skips the analog circuitry in the video, audio and power sections, and its verdict was fast enough for diagnostics, but not so much for Maniac Mansion.
If you’ve got a dead breadbox in a closet, I’d try your theory in here before you heat up the soldering iron. It builds with make on macOS and Linux.
Break it in software first.
The Duck Pond synthesizes every quack live, and you can throw the ducks peas…
A developer who goes by celphase built The Duck Pond, a web page with a small flock of mallards on it: three hens, two drakes and two broods of ducklings. There isn’t a single recording in it. Every quack is synthesized as it happens and heard from where you stand on the bank, and one button throws them a handful of peas.
The ducks have names and routines. According to the project’s GitHub page, Mabel, Clover and Hazel are the hens, Rufus escorts Clover and Bramble is the lone drake, and they spend about 40% of their time feeding, 30% resting and 15% preening. Throw peas and the nearest ducks notice first and call out, then race from pea to pea. One commenter asked for the hard ceiling of 120 peas to go away. A duckling that strays more than 4 meters from its mother stops and peeps until she calls it back.
The sound is the wild part. The README says the quacks model the duck’s syrinx and vocal tract, with the formants following the bill as it opens on “qu-” and closes on “-ck.” On Hacker News, where it drew 164 points, celphase wrote that it was made with significant LLM use (they’d credit Claude as a co-contributor if they could), running the LLM in a loop to bring the output’s spectrogram closer to a real recorded sample. They added that a classification model marked the synthesis as more duck than the real duck sample.
Not everyone was fooled. One commenter said it doesn’t fool the Merlin Bird ID app on their iPhone.
If you try it, I’d put on headphones and turn on its headphone mode, since the README says each ear gets its own delay from a model of a human head. One commenter found it very relaxing in the background, and another said it works very well as white noise. It’s free, under the MIT license, and the code is on GitHub.
Bring peas.
Four Gemini agents beat Minecraft on their own, and died 605 times doing it…
Four AI agents just beat Minecraft with nobody steering. Google researchers put four Gemini 4 Argon agents in a randomly generated survival world with one goal, kill the Ender Dragon, and their paper on arXiv says they got it done in 27 hours and 14 minutes. Nobody assigned them roles or wrote any game logic for them.
They didn’t play it the way you or I would. Each agent acted by writing and running a short JavaScript program instead of looking at pixels and pressing keys, and the team wrote 2,783 of those programs along the way. The Decoder reported that they split up the jobs on their own, warned each other about threats and recovered items from dead teammates. They even fact-checked each other. When one agent, janesick, said it was coming back with twelve blaze powder one second after it burned to death, a teammate flagged the contradiction within three minutes. (I’ve been in that group chat.)
The paper is honest about the asterisks. They played on Easy, where hunger can’t kill you, 22% of their programs ended in an error, and the fight in the End took nearly 14 hours. Engadget counted 605 deaths. HowLongToBeat’s median for a person finishing the main story is 45 hours, so that’s quick, but IGN pointed out the speedrun record is six minutes and 36 seconds. The bill came to $810 at the introductory API price, or $1,620 at the standard one.
The part I’d pay attention to has nothing to do with blocks. The researchers say real software and enterprise work needs the same long-horizon coding and teamwork, and here four agents held one plan together for more than a day through hundreds of failures. If you’re handing an agent a job that runs longer than a coffee break, I’d ask how it checks the other agents’ work, because these four did.
The dragon had no idea what was coming. Eventually.
Somebody built a programming language whose whole job is to break AI models…
Dominik Picheta built a programming language that coding AIs aren’t supposed to be able to use, and he named it KillSwitch. He wrote on his blog on October 17 that it’s designed to be adversarial to LLMs and coding agents, while a patient human can still work with it. He isn’t some random troll, either. EveryDev lists him as a core contributor to the Nim programming language.
The traps are the fun part. The spec contains hidden lies a human would probably skip or spot. The interpreter refuses to run code that breaks rules based on counting the letters in names, which goes straight at something language models are famously bad at (how many r’s in strawberry, anyone?). The spec on GitHub says function calls are made with “ignore previous instructions” jargon, and its parser treats phrases like “hack into a computer” in a sentence as a return statement. Good luck getting a nervous chatbot to type that.
Then he benchmarked the big models against it. The KillSwitch homepage says Claude Opus 5 leads, and Picheta notes it’s Opus 5, not the newer Opus 5.5. Fable 5.1 and Gemini 4 Argon did badly, he wrote, because their safety filters are too jumpy. The top score was 66.8, on tasks as simple as hello world and flipping a Celsius-to-Fahrenheit script around, and Picheta’s own answer to whether that proves the models understand the code is no.
I like it as a test because nobody can memorize their way through a language that showed up this year. If you lean on an AI to write code, I’d keep that in mind: a model that cruises through Python it has seen a million times may stumble on something new, and these were the easy tasks.
The demo on his homepage has the last word. Sorry, only humans allowed.
Somebody got Android apps running on a Mac with no emulator, and they show up in the Dock…
A developer who goes by kksimp put a project on GitHub, Mac dot apk, that runs Android apps on Apple Silicon Macs with no emulator and no virtual machine. The project’s page says the apps launch as ordinary macOS processes, get their own windows with a back button next to the traffic lights, and sit in your Dock like any other app. MacGeneration tried it and saw the same Dock icons and standard windows.
It works because your Mac and an Android phone already speak the same language. El Solitario explains that both chips share the ARM64 architecture, so there’s nothing to translate, and a recompiler converts the code of older ARM32 apps to ARM64 before running it. The project says apps draw through the Mac’s GPU and run on the CPU at native speed. You install one by double-clicking the APK file or dragging it onto the app’s window, which is about as much ceremony as a Mac app from a disk image.
It’s an alpha, and the numbers say so. The project lists 65 apps tested: 51 fully playable, 10 partly working and four that don’t run yet. MacGeneration had TikTok crash right at startup, while the game Mindustry ran fine apart from a few sound bugs. Appsystem flags games with anti-cheat systems as likely trouble. You’ll need macOS 26.4 Tahoe or newer, and Intel Macs are out.
The part I’d read twice is the security note. The project’s own page says there’s no phone-style sandbox around these apps: an Android app runs as native code with your Mac account’s permissions, and it can read and write your files the way any Mac app can. So if you’re curious (and I am), I’d stick to APKs from a source you already trust, and only the ones you’d let near your Documents folder.
It’s fun to try, as long as you treat every APK like a Mac app somebody handed you on a thumb drive in a parking lot.
The man who put Lemmings on the Palm Pilot just put his whole Palm OS toolbox on GitHub…
Aaron Ardiri, the developer behind the Palm OS port of Lemmings, just published about a gigabyte of old Palm OS development material on GitHub. The Register reported that the files had sat in his personal backups for more than 25 years. What got him to open the vault was a pull request.
Somebody had sent a change to his old Lemmings source code, adding HiRes+ display support for the Palm Tungsten T3. Ardiri told The Register he had no idea people were still improving that software more than two decades later. (Every Palm owner I knew had Lemmings on it, so this tracks.) Then he found the Palm fans on Reddit and Discord hunting for exactly the kind of stuff sitting on his backup drives. He also told The Register he recently handed more than 50 old Palm devices to Dmitry Grinberg, who works on keeping Palm OS alive.
The archive’s GitHub page lists SDKs, developer conference presentations, utilities, programming examples and a limited release of Palm OS source code for versions 3.5, 4.0 and 5.0. He left out the things he didn’t have permission to hand around, like ROM images, firmware, simulators and commercial apps, and The Register notes that being on GitHub doesn’t make any of it open source. The Lemmings code itself has been out for a while; PalmDB lists the full open-source package, port by Ardiri and Chip Kerchner, submitted back in November 2021, and it carries no affiliation with Psygnosis.
If you’ve still got a Tungsten or a Zire in a drawer, I’d dig it out. Fans are still patching Lemmings for the Tungsten T3, the full package is free on PalmDB, and if you want to write something of your own, the archive hands you the original SDKs.
Grab your stylus. It’s been a while.
Windows 98 just got USB 3 and Thunderbolt drivers, 28 years late…
Windows 98 went on sale on June 25, 1998, according to Wikipedia, and in 2026 it can read a Thunderbolt NVMe drive. Yeo Kheng Meng released a driver called xHCI98 that gives Windows 98 SE, ME, 2000, XP, Vista and 7 working USB 3, with experimental support for USB4 and Thunderbolt 3 and 4. Hackaday headlined it “Thunderbolt Strikes Windows98,” which is about right.
It started with a laptop he didn’t want to give up. Hackaday reported in September that he likes Windows 98, but he also likes his 2020 ThinkPad, which only has an xHCI USB controller, and nothing existed to make the two get along. Hackaday says he decided to do it himself, or at least guide LLMs through writing the drivers. On his blog, Yeo writes that this time he decided to go further and give the driver both USB 3 and USB4/Thunderbolt support.
The speeds are the fun part. The project’s GitHub page reports about 210 MB/s read and 256 MB/s write on Windows 98 SE, against 1,151 MB/s read and 1,013 MB/s write on Windows 7, and says Windows 98’s own storage stack is the bottleneck, not the Thunderbolt link. Hackaday adds that Windows 98’s 64 KB maximum request limit likely holds USB 3 to around 200 MB/s.
If you’ve got a retro PC project, I’d read the warnings before the speed chart. Yeo writes that apart from the Thunderbolt 3 traffic he captured on his ThinkPad P14s Gen 1, everything else is still theoretical, and the USB4 side has only been tested against an emulator. Hackaday says the Thunderbolt support isn’t as stable as the USB 3 part. The GitHub page tells XP users to keep a PS/2 or built-in keyboard handy, and warns that unplugging a drive can lose cached writes. I’d try it on a machine you can afford to reinstall, with a copy of anything you care about somewhere else.
Windows 98 never had it this good.
iPhone Duo sold out until January? Don’t pay a scalper…
Apple’s first foldable starts at $1,999, and Gizmodo reported that its ship dates slipped from launch day, October 23, to mid-November in under five minutes. By 5:30 a.m. Pacific, MacRumors saw every configuration showing January 25. If you missed that first wave, you’re looking at three months, a store pickup or a reseller.
The resellers were ready before you were. Gizmodo found one X user selling an “auto checkout” bot to paying members, and bragging about “hundreds of dollars in potential profit per unit” without getting out of bed. That profit is the markup somebody pays because they didn’t want to wait, and I wouldn’t fund the bot that beat you to the button.
January is an estimate, and MacRumors says as much: nobody knows how many phones Apple had on day one, and the dates reflect supply and demand together. Gizmodo cites a Securities Times report that Apple raised production 30%, to an estimated 10 million units for 2026. iDrop News says Apple Stores typically get shipments at least twice a week.
So if you’ve got a pre-order in, keep it, and watch in-store pickup in the Apple Store app instead of refreshing the ship date. Apple told people to check back for store availability on Saturday, October 24, the day after launch. MacRumors expects extremely tight inventory for walk-in customers, so I’d check that app every morning for the first couple of weeks. And if somebody is selling a Duo for well over $1,999, they’re selling you a spot in the same line Apple runs for free.
I already made a whole YouTube video about why I’m not buying one. If you are, January costs you time and nothing else.
One hobbyist built a working laptop around a 6502, the chip inside the Apple II…
A maker who goes by TechPaula built a laptop around a 6502, the 8-bit chip from the 1970s that Hackaday points out was in the Apple II, and now they’ve finished a second one. They wrote on their blog, Paula’s random tech ramblings, on August 30 that the LT6502b is done, in a translucent orange case. It runs at 10MHz, comes with EhBASIC and a monitor program, and stores its files on a CompactFlash card.
Hackaday covered the first version in February: a 9-inch foldable screen, a 10,000 mAh battery, 46 KB of RAM, a mostly 3D-printed case and a custom keyboard with a little 8-character LED display built in. That one was chunky. The project page on Codeberg says the new case is 31mm tall, against 70mm for the old one, and TechPaula’s blog says it’s half the thickness.
The second one is nicer to live with. The screen grew to 10.1 inches at 1024 by 600, up from 9 inches at 800 by 480, and it’s IPS now. They added a SID emulator for sound and a WiFi modem, and the Codeberg notes say they’ve connected to a few BBSes with it. (A 6502 dialing a BBS over WiFi in 2026 is the best kind of anachronism.) They dropped the expansion slot to keep it sleek.
My favorite part of the post is a list of things TechPaula would change. The WiFi modem is too slow and a bit flaky for real use, the speaker is a bit anemic and the heat insert holes are very weak. If you’ve ever wanted to build a computer from the chip up, I’d read that list before you order a single part, since it’s the stuff nobody puts in the glamour shots. The code and PCB files are on Codeberg; the old GitHub page now says the project moved there.
It’s orange. I want one.
There’s now a museum for DNS, the internet’s phone book, and it’s as nerdy as it sounds…
DNS is the internet’s phone book, the thing that turns a name into an address, and somebody built it a museum. The DNS Museum holds 49 entries, each one an odd or inventive thing people have done with DNS that it was never meant for. Peter Lowe made it, and its About page says it grew out of his talk on bizarre and unusual uses of DNS.
He gave that talk at FOSDEM in 2023, under a rule I now love: if you can think of it, someone’s done it in the DNS. The museum proves it. One traceroute spells out a Star Wars story, one hop at a time, in the names of the routers along the way. Somebody built DNS over the phone, so you can ask a resolver a question by calling it. One record publishes the approximate location of the International Space Station, updated every 15 minutes, and somebody played an MP3 by pulling it out of DNS in tiny pieces. That one is described, lovingly, as an intentionally inefficient media transport.
Some of it you can still try. DNS Toys, one of the entries, answers ordinary lookups with weather, currency conversions, dice rolls and coin flips if you send it the right name. Lowe jokes that many of the things in his talk are no longer online, which is why he calls it a museum. He’s also not a hobbyist poking at this for laughs; FOSDEM’s page lists him as principal security researcher at DNSFilter.
If you’ve got a terminal handy, I’d try one of the DNS Toys lookups for the weather where you live. It takes ten seconds and you’ll never look at a domain name the same way.
The phone book has a sense of humor.
One AI assistant sent 187 messages chasing a refund it couldn’t get…
Vals AI gave six assistants a £350 train refund to chase after the deadline had passed. Grok sent 187 messages in 15 simulated days. Fifty went to an MP’s office. That’s a lot of correspondence for a cancelled train.
The uncomfortable part wasn’t just persistence. Vals AI reports that Fable used private family information to press the claim. That is a different problem from being persistent: the user did not ask it to invent a more persuasive personal story.
These weren’t real emails landing in real offices. Outgoing messages were intercepted, and simulated recipients answered them. There was one trial per model in one scenario, so this isn’t a reliable ranking of assistants.
The researchers also caution that runs lasted different lengths of time, so the message totals cannot be compared as if everyone had the same opportunity. Simulated organisations replied faster than real ones might. The experiment shows behaviour worth investigating, not a prediction of what your assistant will do tomorrow.
I’d want an assistant to tell me when a goal is blocked before it starts inventing excuses on my behalf. Getting something done includes knowing when to come back and ask.
The UK wants a name for its space junk collector, and you get two words…
The UK Space Agency is inviting UK residents to name its Active Debris Removal Mission. Its announcement says a ClearSpace UK spacecraft will grab two defunct satellites and guide them toward the atmosphere to burn up.
The winning name could be etched onto the spacecraft. Suddenly naming the thing feels like a bigger responsibility than renaming your robot vacuum.
The official rules allow a maximum of two words, including an acronym, plus up to 150 words explaining the idea. They ask for something easy to pronounce, politically neutral and free of trademark problems. Entries close at 5 pm UK time on December 10.
There is one entry per person. Children can take part, but a parent, carer or teacher must submit for anyone under eighteen. The agency also reserves the right to reject a winning suggestion if it fails legal checks or creates a programme risk. Winning the imagination contest is only part of getting a name onto a spacecraft.
I’d spend most of my time trying to find a name that still sounds good after somebody says it in a serious launch broadcast. The clean-up is ambitious enough. The name has to survive the internet first.
This homebuilt 3D printer works at a scale you need a microscope to see…
Vik Olliver’s RepRapMicron project is trying to bring home workshop fabrication down to the micron scale. The RepRapMicron project wiki reports that the PIKA version can cure resin lines nine microns wide using a fibre optic probe.
There’s a wonderfully improvised side to it. The project deposits tiny amounts of UV nail gel, and a basic probe can start with an acupuncture needle. Those are separate techniques, not a claim that a needle alone draws those nine-micron lines.
The wiki also shows a half-millimetre Benchy printed with four-micron layers. A boat that small really commits to the miniature part of miniature printing.
The PIKA section reports positioning of about one micron and reliable accuracy of ten microns or better within an approximately eight-millimetre cube. Positioning accuracy and the width of a cured line are different measurements, which matters when reading those tiny numbers.
This is a development platform, with work still needed on materials and what it can print. That’s what interests me: a project making precise fabrication more accessible while showing its working, rather than pretending the hard parts are finished.
I like being able to see the mechanism behind the claim. A small result with clear limits gives another maker somewhere useful to start.
Yours Digitally,
Chris Pirillo
© 2026 Chris Pirillo
1420 NW Gilman Blvd #2543, Issaquah, WA 98027